Questions/Comments?
Mail me
|
|
| GOAL |
| To represent a 3D object
with a minimal set of 2D views sufficiently enough to recognize the object
amongst a set of others from an unknown view. |
Figure 1. The set of models used in recognition experiments
| OVERVIEW OF PROCEDURE |
| The goal of the aspect-graph representation
is to partition the viewing space into a minimal set of views that can
be distinguished as a group to determine view transitions, corresponding
to visual events, e.g., as a new part comes into view. Since traditional
methods based on the singularities of visual mapping are not applicable
to complex free-form objects and also often result in numerous aspects,
we adopt an approach based on grouping views into aspects using a notion
of similarity between views. One can abstractly view the similiarity-based
aspect generation approach as performing "edge detection" on the viewing
sphere by analyzing projections of the 3D object. In contrast, the
aspect generation method of using similarity of adjacent views can be viewed
as a "region-growing" segmentation approach. This has two distinct
advantages. First, the salience of a singularity in the visual mapping
is related not only to its own significance but also on the lack of such
events in its neighboring views. Second, the grouping of similar
views can be done exclusively in the domain of 2D images without requiring
or restricting 3D representations of shape. |
|
